For most companies, building an AI prototype is no longer the difficult part. A capable team can get a model answering questions about internal data in a few weeks. The harder questions come weeks later, however.
How much does each interaction cost when thousands of customers use it everyday? Who notices when it starts giving bad answers? Which billing system or compliance process breaks when it changes something upstream? And who is responsible for it after the consultants leave?
This helps to explain why so much enterprise AI never makes it past the pilot stage. An MIT NANDA study published last year found that 95% of corporate generative AI pilots delivered little to no measurable impact on profit and loss, and Gartner projects more than 40% of agentic AI projects to be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls.
Such a wide gap between demo and production is now shaping how AI model providers decide which partners to vouch for. This month, for instance, Anthropic named Making Sense, a technology consulting firm with more than 20 years of software work for mid-market and private equity-backed U.S. companies, a Select partner in the Services Track of its Claude Partner Network.
The tier is awarded on the basis of systems a firm has put into production and that clients run every day, not on pitch decks or announced alliances.
A credential based on delivery
Companies shopping for AI help face a sorting problem: “agent washing”, which Gartner describes as a term for vendors rebranding chatbots and existing automation tools as agentic AI.
The consulting firm estimates that only about 130 of the thousands of vendors claiming agentic capabilities actually offer them.
Anthropic’s own partner program drew more than 40,000 applications within months of its March launch, which was backed by a $100 million commitment to partner training, technical support and marketing.
The Services Track, introduced in June, is Anthropic’s attempt to separate firms that have shipped from firms that have applied. Each tier depends on three things a firm can show: practitioners holding current Anthropic certifications, customers taken live with Claude in the trailing 12 months, and public customer stories.
The published thresholds for Select start at 10 certified practitioners and two production customers; Making Sense reports 20 certified staff.
Fernando Florez, the firm’s chief revenue officer, sees the value less in the badge than in what a buyer can actually verify. “Anthropic reviews each partner’s standing every quarter, based on its certified professionals and the customers it has taken live with Claude,” he said.
“That gives clients a public, verifiable reference for what we’ve delivered, which makes choosing who to work with less of a leap of faith.”
Firms in the program are listed in a public Claude Partner Hub directory, where companies can compare tier, team certifications, deployments and references side by side.
What production actually demands
The project behind Making Sense’s designation shows how much of AI in production comes down to unglamorous work. Doppler, one of Latin America’s most established digital marketing platforms, runs email and marketing automation for clients across Latin America and Spain, and Making Sense built it into an assistant called ECO IA.
The assistant lives inside the product, pulls live account data, checks how healthy an account is and writes draft campaigns. Users then review and edit those drafts before anything goes out.
Keeping a human in that review step is a deliberate design choice for a tool that touches customer communications. The assistant runs on Claude Sonnet 4.5 and is enabled by default on every new Doppler account, so it had to work for thousands of users from the start rather than for a curated test group.
According to Making Sense, it is now available in more than 4,800 accounts, and 95% of users who engaged with it came back across multiple months.
Cost is where many AI products fail after launch, and the economics of ECO IA depend on prompt caching, which lets an application reuse context it sends repeatedly instead of paying to process it again on every call.
The firm reports a 97% cache hit rate and an inference cost of $0.066 per interaction, and it also built a real-time dashboard tracking usage, cost, latency and errors—turning an AI feature into something an operations team can manage and budget for explicitly.
Who answers for it once it’s live
Integration was the main reason MIT’s researchers gave for stalled pilots. The study’s lead author, Aditya Challapally, told Computing that enterprise AI depends on careful integration with existing systems and that pilots rarely scale without it.
César D’Onofrio, CEO and cofounder of Making Sense, argues that faster model-driven coding has made this kind of institutional knowledge more valuable. “Claude has made building software dramatically faster. Knowing what a change will touch inside a company still takes people who understand that business: the system billing depends on, the data a regulator will ask about, the team that has to operate it the next morning,” he said.
“We’ve been doing that work for more than 20 years. We decide with our clients what is worth building, connect it to everything their operation runs on, keep it compliant and reliable, and answer for it once it’s live.”
Part of answering for a system is ensuring the client can run it without outside help. Making Sense often embeds its engineers with the client’s own developers; at Doppler, the engagement included automated testing and workflow changes that left the in-house team able to keep validating and extending its Claude integration.
At Viallion, meanwhile, an AI-powered trading automation platform, the firm rebuilt parts of the engineering workflow around Claude Code. The company says initiatives that once took two to three months to develop, validate and release now ship in about a week, and that Viallion’s engineers have kept refining the process on their own since the enablement phase ended.
A large portion of Making Sense’s clientele sits in healthcare, fintech and logistics, and other sectors where a regulator, an auditor, or a payments processor will eventually ask how an AI system reached its output.
Here especially, the firm that built the system needs to be able to answer the questions that come after launch.
Featured image: Alex Shuper via Unsplash+
Disclosure: This article mentions a client of an Espacio portfolio company.
